What they did
Roledene built a small event ticketing demo app (FastAPI, SQLite, HTMX, pytest) with Claude Code set up as a team. CLAUDE.md holds the shared rules, including which files each role may edit. Seven skills cover the stages from idea to spec, design, tasks, code, tests and review. Six subagents do the work: a spec writer and a designer, then a backend coder, a UI coder and a tester running in parallel against the same design contract, and a reviewer that checks every numbered acceptance criterion against a test and a code location. The tester writes tests from the spec without reading the new code. One command runs the whole pipeline, and failing criteria go back to the agent that owns them, for at most two rounds.
How it went
For the waitlist feature, the only input was a three-line idea. The pipeline ended with all 12 acceptance criteria passing, 114 tests green and clean lint. Roledene stepped in twice: to approve the spec and to skim the design. On an earlier feature the independent tests caught a date handling bug at once. On another, all 53 tests passed but the reviewer spotted that HTMX 2 silently drops 4xx responses, so a cancel button did nothing; one more round fixed it. An agent also ran a formatter over files outside its lane.
Worth knowing
Telling an agent which files it owns was not enough; Roledene had to exclude Markdown from the formatter to stop it. Each parallel agent used about 50k tokens per feature, plus the reviewer, so only truly independent lanes are worth splitting. This is a demo project built for a talk, not a production app.